New preprint 💥 Can a general-purpose model achieve results comparable to medically pre-trained models? 🤔 We show that lightweight fine-tuning of a general-purpose LVLM and an LVLM-aware retriever can. 🚀 🔗 GitHub: github.com/Nirmaz/CLARE 📄 Paper: arxiv.org/pdf/2508.17394

I'll be at #COLM2026 in SF presenting this: Tue, 4:30 PM, poster #95 in Franciscan C. Come say hi! TL;DR: fine-tuning on new facts makes LLMs forget old facts that are semantically similar. Overlap predicts how much, and self-distillation prevents it. What's new in the CR 🧵👇

Camera ready is ready! 📸🥳 arxiv.org/pdf/2601.18987 @emnlpmeeting #EMNLP2026 w/ @jordiae Pascal Kesseli @jvanegue @HyadataLab @adiyossLC @PeterOHearn12

New paper: We’re building retrieval models for AI scientists (and human ones, too)—retrieval trained to directly support inspiration💡: 🫠Problem-solution latent links, ⏫Zooming out (abstraction), ⏬Zooming in (specification). Cognitive theories of ideation suggests these are fundamental ways humans traverse idea spaces and reframe problems, while brainstorming and connecting different areas. Our new work, RATIO, with my student @maayansharon10, brings these operations into scientific retrieval. Given for example a query on “reward hacking in RL-trained LLMs,” the goal is to retrieve a direct solution (e.g., reward-model ensembles), broaden the problem (e.g., to Goodhart’s law), or specify that principle in a new setting (e.g., recommender systems). Our focus here is on defining the task and creating a new automatic way for data collection from literature. The overarching goal: retrieve not just according to standard notions of topical relevance, but according to more latent, abstract links that help scientists take the next ideation step. I've been exploring this line of work in multiple papers over the years, excited to take a first step toward training models for this, and for more upcoming work in this space. This is also related to the recent notion of oblique retrieval (@lateinteraction): retrieving latent and implicit connections rather than relying only on surface similarity.

Mark your calendars for ISCOL '26! 🤩 We're excited to announce that this year's ISCOL conference will take place on December 23rd at The Hebrew University. See you soon in Jerusalem:)

Manipulation leaves fingerprints. 4 principles for investigating disinformation before you call anything a psyop: 01 Hunt the trail, not the story 02 Ask what it was built to do to you 03 Three searches are not a verdict 04 Show your receipts I don't hunt chaos. I haunt it.

Did you know you that every Wikipedia article's full edit history is open to the public? What if every scientist, journalist or student could easily explore how any scientific Wikipedia article evolved along the years? Now they can. Meet our WikiSTAR system! 🧵👇🏽

Looking forward to presenting this work at ACL 2026 as an oral presentation! 🗓️ Sunday, July 5 🎙️ Oral Session D: Special Theme – Explainability of NLP Models 1 🕘 9:30 AM 📍 Harbor G Come by if you’re attending ACL! happy to chat about the paper!

At #ACL2026 with two papers! 🥗 Vocab Diet, led by @YuvalReif: July 6, 11:00, session E 🌊 Follow the Flow, equal cont @michael_toker: July 7, 11:00, session G I study how LLMs organize meaning: tokenization, interpretability, efficiency. Come say hi! 👋 Happy 4th of July! 🇺🇸🇺🇸🇺🇸

If you're attending #ACL2026 later today, check out these talks from my students and colleagues (along with all the other cool work from @allen_ai )! @NoySternlicht @NirMMazor Lior Forer Hiba Arnaout @IGurevych @UKPLab @aclmeeting See details in the image.

💡As #ACL2026 kicks off, we have four papers presented by my students and collaborators on using AI and scientific literature for discovery. See details in the image ⬇️ and links in the thread 🧵 @NoySternlicht @NirMMazor Lior Forer Hiba Arnaout @IGurevych @allen_ai @nlphuji @aclmeeting @UKPLab

🎉 Happy to share that CHIMERA has been accepted to #ACL2026NLP (main conference)! 📄 Paper - arxiv.org/abs/2505.20779 🤗 Data - tinyurl.com/yhncsu34 🌐 Project Page - noy-sternlicht.github.io/CHI… 💻 Git - github.com/noy-sternlicht/CH… Joint work with @Hoper_Tom

New datasets keep coming, New models keep coming. Frustrating! How can we evaluate everything on everything? How do we keep scores comparable over time? We propose a way to grow benchmark suites without losing comparability. Details:👇🧵

Have LLMs become supervised learners (once again)?! In our new paper, we argue that current LLMs’ post-training methods have effectively reverted to the "pre-train then fine-tune" era, explicitly tailoring models to desired behaviors. 1/n

Fine-Tuning LLMs on New Knowledge Encourages Hallucinations. (@zorikgekhman) But why? We found something unexpected: 1M facts about city-like names →hallucinations explode. 1M facts about random identifiers →near zero! Same model. Same number of facts. Only the names change.🧵

Ever used a top-ranked LLM that just... felt wrong for you? You’re not alone. Instead of leaderboards, many of us turn to "vibe-testing" - manually comparing models to our own needs. But can we turn these feelings into a structured evaluation? New paper: "From Feelings to Metrics" 🧵

We introduce 🌍GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens.🌍 Most feed-forward 3DGS methods still start from pixel, voxel, or dense view-aligned primitives. We take a different route: align first, decode later. 🧵👇

We are excited to share our new preprint! 🚨 biorxiv.org/content/10.1101/… In this work, we investigate the neural mechanisms by which speakers generate information-rich linguistic content during spontaneous natural conversation.

What if you could automatically turn a large collection of documents into structured databases, tailored for your own research needs?📄 We introduce ScheMatiQ! From question ➝ schema ➝ structured data 🔍 @ShaharLevy19 @MintzReshef @RKeydar @BarakRaveh @GabiStanovsky 🧵 👇1/5

Happy to share three papers accepted to the #ACL2026 @aclmeeting conference: on AI ideation, multimodal medical AI, and literature understanding. If you want to read about them, check out the following posts from @NoySternlicht , @NirMMazor and @UKPLab . More updates on each to come separately. In the hypothesis generation / ideation space : cross-domain inspirations. Led by @NoySternlicht . nitter.siddz.com/NoySternlicht/status/1… In the multimodal medical AI domain: medical image diagnosis with literature retrieval. Led by @NirMMazor . nitter.siddz.com/NirMMazor/status/20157… In literature understanding space: summarizing thousands of citations to understand a paper's impact. Led by Hiba Arnaout from @IGurevych's lab, with major contributions from @NoySternlicht . lnkd.in/de6DCS8J

Today at #EACL2026 we present our work exploring PRINTED newspapers as a data source for summarization in low-resource languages. Join the virtual session at 18:00! 🧵👇